Rats are social animals. For example, rats exhibit mutual-reward preferences, preferring choice alternatives that yield a reward to themselves as well as to a conspecific, over alternatives that yield a reward only to themselves. We have recently hypothesized that such mutual-reward preferences might be the result of reinforcing properties of ultrasonic vocalizations (USVs) emitted by the conspecifics. USVs in rats serve as situation-dependent socio-affective signals with important communicative functions. To test this possibility, here, we trained rats to enter one of two compartments in a T-maze setting. Entering either compartment yielded identical food rewards as well as playback of pre-recorded USVs either in the 50-kHz range, which we expected to be appetitive or therefore a potential positive reinforcer, or in the 22-kHz range predicted to be aversive and therefore a potential negative reinforcer. In three separate experimental conditions, rats chose between compartments yielding either 50-kHz USVs versus a non-ultrasonic control stimulus (condition 1), 22-kHz USVs versus a non-ultrasonic control stimulus (condition 2), or 50-kHz versus 22-kHz USVs (condition 3). Results show that rats exhibit a transient preference for the 50-kHz USV playback over non-ultrasonic control stimuli, as well as an initial avoidance of 22-kHz USV relative to non-ultrasonic control stimuli on trend-level. As rats progressed within session through trials, and across sessions, these preferences diminished, in line with previous findings. These results support our hypothesis that USVs have transiently motivating reinforcing properties, putatively acquired through association processes, but also highlight that these motivating properties are context-dependent and modulatory, and might not act as primary reinforcers when presented in isolation. We conclude this article with a second part on a multilevel cognitive theory of rats’ action and action learning. The “cascade” approach assumes that rats’ cognitive representations of action may be multilevel. A basic physical level of action may be invested with higher levels of action that integrate emotional, motivational, and social significance. Learning in an experiment consists in the cognitive formation of multilevel action representations. Social action and interaction in particular are proposed to be cognitively modeled as multilevel. Our results have implications for understanding the structure of social cognition, and social learning, in animals and humans.